NVIDIA's upcoming Vera data center CPU brings serious competition to the x86 landscape with its custom Olympus cores. The ARM-based processor packs 88 cores targeting agentic AI workloads, delivering performance that rivals Intel and AMD offerings in ways rarely seen from non-x86 chips. Early testing reveals promising results ahead of the second half 2025 launch.

Each Olympus core supports Armv9.2 instructions and FP8 precision while achieving 176 threads through spatial multi-threading. The design pairs with LPDDR5X memory for up to 1.2TB/s bandwidth, includes 2MB L2 cache per core, and features a massive 164MB unified L3 cache. Connectivity includes PCIe Gen 6 and CXL 3.1 support, positioning Vera as a high-bandwidth solution for modern AI infrastructure.

Linux support appears well-developed despite the pre-release status. The processor works with kernel 7.1+ and distributions like Ubuntu and Fedora, thanks to Arm SBSA compliance. Compiler support arrived early through GCC 16.1 and LLVM Clang 21, beating typical ARM timelines and matching Intel's proactive approach to developer tools.

Testing occurred on pre-production hardware with 450 Watt TDP and some power management features still in development. NVIDIA restricted benchmarks to specific workloads relevant to data center customers, leaving broader performance questions for later evaluation. The early results suggest ARM may finally challenge x86 dominance in AI compute.